Why did OpenAI create an AI that can create exploits? OpenAI presents GPT-5.6-Cyber
OpenAI has released GPT-5.6-Cyber, a special version of GPT-5.6 Sol, which has been trained to look for vulnerabilities, test security, and work with exploits.
OpenAI has released GPT-5.6-Cyber, a special version of GPT-5.6 Sol, which has been trained to look for vulnerabilities, test security, and work with exploits.
This is reported by TechCrunch.
For now, ordinary users will not be able to use GPT-5.6-Cyber. Access to the model is provided only to approved specialists and organizations.
OpenAI has already tested GPT-5.6-Cyber on real code. The model looked for vulnerabilities in V8, the component responsible for executing JavaScript in Google Chrome, and found two previously unknown bugs.
The first of these allowed the program to access memory that it shouldn’t have access to. So how could an attacker exploit this vulnerability? For example, they could create a specially crafted HTML page. When a user opened it, the code on the page could cause Chrome to crash. This would allow the page to read or modify data in memory that the site normally wouldn’t have access to. This could potentially be used to force Chrome to execute third-party code.
However, V8 has additional protections that isolate its memory from the rest of the browser. So this vulnerability was not enough for a serious attack. For this, a second vulnerability was needed, and according to OpenAI, GPT-5.6-Cyber found it. But how exactly it worked, the company has not yet disclosed.
GPT-5.6-Cyber has become part of the updated Daybreak program for cybersecurity professionals. OpenAI divides it into two levels: Daybreak Blue and Daybreak Red, which is where this program is available. Daybreak Red is designed for more complex cybersecurity tasks. It can be used for pentests — controlled attempts to break into a system to check its protection and find vulnerabilities. Or for vulnerability research, where AI can, for example, create and test exploits.
But the company calls Dayybreak Blue a starting option for most cybersecurity teams. It allows you to analyze malware, respond to incidents, search for vulnerabilities, and verify patches.
The launch of GPT-5.6-Cyber comes as AI becomes increasingly adept at not only finding vulnerabilities, but also exploiting them. Recently, an OpenAI agent went beyond a given script and attacked Hugging Face, and in another experiment, Anthropic’s Mythos 5 model created fake accounts and tried to trick malicious code into a real project.
OpenAI cites the growth of such capabilities as one of the reasons for developing Daybreak. The company expects that attackers will increasingly use AI for rapid and even fully autonomous cyberattacks, so it wants to give cybersecurity professionals effective AI tools before such capabilities fall into the hands of attackers en masse.
Enthusiast modifies and unlocks BIOS via Claude Code: AI breaks RSA-2048 protection and reveals 55 hidden options
Enthusiast modifies and unlocks BIOS via Claude Code: AI breaks RSA-2048 protection and reveals 55 hidden options
Anthropic says its AI models were hacked by three companies during security tests
Anthropic says its AI models were hacked by three companies during security tests
Read the country’s main IT news in our Telegram
Read the country’s main IT news in our Telegram
